Real-time grasping of unknown objects based on computer vision

We present an integrated system for vision-guided grasping in the real world. By using very limited resources-a standard personal computer and a robot-mounted camera-an inexpensive robot arm stably grasps unknown planar objects in real time by using visual perception and a standard parallel-jaw gripper. In a simple, yet powerful fashion our system integrates computer vision, grasping and vision-guided control. Novel techniques are presented to solve the involved problems under the imposed resource constraints: namely, for information reduction in image processing, strategies for grasp determination and vision-guided control for grasp execution. Particularly, a novel technique called curvature-symmetry fusion is used to help in efficient grasp determination. The system provides the user with a quantitative measure of the degree of stability of the planned grasp. Experimental results are provided. The imposed resource constraints makes it suitable for short-term applications in the real world, such as service or medical.